Skip to content

futureagi.md

Latest commit

 

History

History
56 lines (41 loc) · 1.89 KB

futureagi.md

File metadata and controls

56 lines (41 loc) · 1.89 KB
Raw

Future AGI evaluation

CortexFS keeps agent execution provider-neutral and records completed sessions as ATIF trajectories. The optional cortexfs-futureagi adapter connects that existing evidence to Future AGI without adding a watcher, a background uploader, or a second /ctx ABI.

Build

cargo build --release -p cortexfs-futureagi

The binary is target/release/cortexfs-futureagi. It is intentionally an optional adapter rather than part of the runtime crate's provider path.

Export a trajectory

First produce a validated ATIF projection using the existing command:

ctx agent trajectory executor --session default > trajectory.json
cortexfs-futureagi export --trajectory trajectory.json > futureagi-data.json

The exported array joins user inputs and includes the last agent output. --include-context additionally includes tool observations as context; leave it off when those observations contain private data.

For local metrics, pass the exported JSON directly to the upstream CLI:

fi run --mode local --eval answer_relevancy --data futureagi-data.json

Cloud evaluation

The command sends the trajectory's user and agent text to the configured endpoint. Review the exported data and your data-handling policy first.

The cloud adapter reads FI_API_KEY, FI_SECRET_KEY, and optionally FI_BASE_URL from the environment. It submits one trajectory case to the Future AGI evaluation API and prints the JSON response:

export FI_API_KEY=...
export FI_SECRET_KEY=...
cortexfs-futureagi evaluate \
  --trajectory trajectory.json \
  --eval answer_relevancy

Use --base-url for a compatible self-hosted endpoint and --timeout to change the default 200-second request limit. Keys are never written to session files. Evaluation is an explicit command, so Git commit remains the only runtime activation boundary.